In 2026, developers bypassed the software restrictions of the Apple M4 chip's neural engine through reverse engineering, unlocking its hidden AI training capabilities. This breakthrough leverages a custom Model Intermediate Language (MIL) to communicate directly with the chip, bypassing dependencies on CoreML, Metal, and GPU, enabling the M4 to achieve 15.8TFLOPS of AI computing power on consumer devices and support Transformer model training. The core of the technology lies in running the entire data workflow in RAM to avoid performance bottlenecks and invoking low-level instruction sets through a custom toolchain. Experiments show that the M4 achieves an energy efficiency of 6.6TFLOPS/W when training a single-layer Transformer model, with power consumption below 1 watt. However, limited by memory bandwidth and hardware locks, the current implementation still faces engineering challenges.